Urban air quality proxy mapping from satellite aerosol and NO2 data
Satellite-derived aerosol optical depth and tropospheric NO2 columns can be disaggregated to sub-kilometre resolution using land-use regression, producing spatially continuous pollution proxies where ground networks are absent or patchy.
Sensors
- Sentinel-5P TROPOMI: Measures tropospheric NO2 columns at a native pixel size of 3.5 × 5.5 km (upgraded from 7 × 3.5 km in August 2019). Daily global coverage. The coarse footprint is the central resolution constraint for urban disaggregation work; individual street canyons are invisible to it directly.
- MODIS MAIAC AOD: Multi-Angle Implementation of Atmospheric Correction algorithm retrieves aerosol optical depth at 1 km resolution from Terra and Aqua MODIS, with one to two overpasses per day per platform. MAIAC outperforms the standard Dark Target product in urban areas but remains blind under cloud and confused by bright desert dust unrelated to combustion emissions.
- Sentinel-2 MSI: 10 m multispectral imagery used as a land-use regression input: road density, building density, green fraction and impervious surface fraction are derived from Sentinel-2 classifications to spatially distribute coarse satellite pollution signals. 5-day revisit at the equator with two satellites.
- GEMS (Geostationary Environment Monitoring Spectrometer): South Korea's GEMS, operational since 2020, measures NO2, SO2, O3 and aerosols over East and Southeast Asia at roughly 3.5 × 8 km, with hourly daytime observations. Its geostationary position allows diurnal pollution cycle tracking that polar-orbiting sensors cannot provide.
What a 3.5 km pixel actually contains
TROPOMI's tropospheric NO2 column is one of the most scientifically mature satellite air-quality products available. It is also, for city planners, frustratingly coarse. A single 3.5 × 5.5 km pixel over a typical Asian or African city contains a motorway interchange, a residential neighbourhood, a market and a park. The pixel average tells you the city has a problem. It does not tell you which block to regulate.
This is not a calibration failure or a processing shortcut. It is a physical limit set by the signal-to-noise requirements of UV-visible spectrometry from 824 km altitude. Sentinel-5P cannot be sharpened by better algorithms alone. The disaggregation step, using land-use regression, is therefore not optional decoration; it is the analytical core of any sub-kilometre urban product derived from TROPOMI.
How land-use regression bridges the resolution gap
Land-use regression (LUR) was developed originally for ground-monitor networks: you build a statistical model that predicts pollution at unmonitored locations from observable spatial predictors such as road length within a buffer, traffic counts, distance to industrial sites and population density. The same logic applies when the dependent variable is a satellite column rather than a ground measurement.
For satellite-based LUR, Sentinel-2 classifications supply the spatial predictors at 10 m. Road network density, impervious fraction, building footprint density and proximity to industrial land parcels are computed at multiple buffer radii (typically 100 m, 300 m, 1 km). A regression model, trained where ground monitors do exist, redistributes the TROPOMI NO2 column across the finer grid. Published studies using this approach in cities with dense monitor networks report cross-validated R² values in the range of 0.6 to 0.8, though performance degrades in cities where the training data are sparse or where emission sources are highly localised.
The honest caveat: the model inherits whatever biases exist in TROPOMI's cloud-screened retrievals and in the land-use classification. A misclassified industrial zone propagates directly into the disaggregated surface.
Aerosol optical depth as a PM2.5 proxy, and where it breaks down
MODIS MAIAC AOD at 1 km is widely used as a proxy for surface PM2.5 concentrations. The physical basis is real: combustion aerosols scatter and absorb sunlight, and the column-integrated optical depth correlates with surface particle loading under stable boundary-layer conditions. Published conversion factors (aerosol extinction efficiency, boundary-layer height corrections) allow approximate PM2.5 estimates from AOD, and the approach is documented in peer-reviewed literature going back to the early 2000s.
The confounders are serious and must be named. Dust events, common across the Sahel, Arabian Peninsula, South Asia and northern China, inflate AOD dramatically without any corresponding increase in urban combustion emissions. A Saharan dust intrusion over a West African city can push MAIAC AOD above 1.0 while ground-level PM2.5 from traffic remains moderate. Separating dust from urban combustion aerosol requires ancillary data: aerosol absorption optical depth, Ångström exponents from AERONET stations, or back-trajectory modelling. Without that separation, PM2.5 estimates from AOD alone are unreliable during dust episodes.
Cloud cover is the other hard limit. MAIAC cannot retrieve AOD through cloud. In tropical cities with persistent convective cloud cover, data gaps can run to weeks at a time. Monthly composites reduce the gap problem but lose the temporal resolution needed to track acute pollution events.
GEMS and the diurnal dimension
Polar-orbiting sensors like Sentinel-5P and MODIS observe each city at roughly the same local time each day. TROPOMI crosses the equator at around 13:30 local solar time. Morning rush-hour NO2 peaks, which often exceed the afternoon observation by a factor of two or more in cities with heavy commuter traffic, are invisible to it.
GEMS changes this for its coverage domain. Hourly NO2 retrievals over East and Southeast Asia allow the diurnal pollution cycle to be reconstructed: morning build-up, midday photochemical breakdown, afternoon secondary formation. For health-impact assessments that need to account for population exposure during peak commuting hours, this temporal dimension matters. The trade-off is spatial: GEMS pixels are larger than TROPOMI's in the across-track direction, and coverage is limited to the Asia-Pacific sector.
What the output is actually good for, and what it is not
Satellite-derived pollution proxies at sub-kilometre resolution are genuinely useful for identifying relative spatial patterns within a city: which districts consistently show higher NO2 or AOD, which industrial corridors stand out, how pollution gradients track road networks. This is enough to prioritise ground-monitor placement, support epidemiological exposure modelling at neighbourhood scale, and flag hotspots for regulatory attention.
They are not a substitute for a calibrated ground network when absolute concentration values are needed for regulatory compliance or health-standard enforcement. The satellite-derived surface is a proxy, not a measurement. Uncertainty bounds on disaggregated NO2 surfaces are rarely better than ±20 to 30 percent in cities without dense monitor networks for model training. Any health-impact assessment built on these products should propagate that uncertainty explicitly rather than treating the gridded output as ground truth.
Satellize runs operational pipelines on TROPOMI and MAIAC data, combining open-constellation inputs with Sentinel-2 land-use regression layers.
Getting to a deliverable: from column to map to decision
A practical urban air-quality proxy product involves three stages. First, multi-year TROPOMI NO2 and MAIAC AOD archives are filtered for cloud fraction below 0.3 and, for AOD, for dust-episode days flagged by ancillary aerosol-type data. Seasonal and annual composites are built to reduce retrieval noise. Second, Sentinel-2 land-use predictors are computed at the target city extent and the LUR model is fitted, with cross-validation against whatever ground monitors exist. Third, the disaggregated surfaces are delivered as georeferenced GeoTIFF or vector polygon layers, with per-pixel uncertainty estimates attached.
Archive depth for TROPOMI runs from May 2018 to present. MODIS MAIAC extends to 2000, giving more than two decades of AOD for trend analysis. That historical depth is one of the most underused assets in urban air-quality work: it allows cities to quantify whether pollution has improved or worsened over the period of specific policy interventions, with no reliance on historical ground-monitor records that may not exist.
Typical figures
| TROPOMI NO2 native pixel size | 3.5 × 5.5 km (post-August 2019 upgrade from 7 × 3.5 km) |
| MODIS MAIAC AOD resolution | 1 km |
| Disaggregated output resolution (LUR) | 100 m to 500 m, constrained by Sentinel-2 predictor resolution and model training density |
| TROPOMI revisit | Daily global coverage; single overpass at ~13:30 local solar time |
| MODIS AOD revisit | 1 to 2 overpasses per day (Terra + Aqua combined); cloud-free retrieval rate varies by region |
| GEMS temporal resolution | Hourly daytime observations over East and Southeast Asia |
| TROPOMI archive depth | May 2018 to present |
| MODIS MAIAC archive depth | 2000 to present (Terra); 2002 to present (Aqua) |
| Typical LUR cross-validated accuracy | R² 0.6 to 0.8 in cities with adequate ground-monitor training data; degrades significantly where monitors are sparse |
| Delivery formats | GeoTIFF (per-pixel value and uncertainty), GeoPackage polygon layers, CSV summary statistics by administrative unit |
Analytics Satellize can run
| Annual mean NO2 concentration proxy surface | TROPOMI tropospheric NO2 column compositing with LUR disaggregation using Sentinel-2 land-use predictors | GeoTIFF at 100–500 m resolution with per-pixel uncertainty band; suitable for epidemiological exposure modelling |
| PM2.5 proxy surface from MAIAC AOD | MAIAC AOD to surface PM2.5 conversion using boundary-layer height correction and dust-episode filtering; seasonal composites | Seasonal and annual raster layers with dust-episode exclusion flags; GeoTIFF |
| Pollution hotspot polygon layer | Spatial clustering (DBSCAN or kernel density) applied to disaggregated NO2 and AOD surfaces to identify persistent high-concentration zones | GeoPackage polygon layer ranked by mean concentration and persistence; suitable for regulatory prioritisation |
| Multi-year trend analysis by district | Mann-Kendall trend test on annual TROPOMI NO2 and MAIAC AOD composites aggregated to administrative boundaries | Tabular report with trend direction, magnitude and statistical significance per district; PDF and CSV |
| Ground-monitor siting recommendation | Spatial variance analysis of disaggregated NO2 surface to identify locations that maximise representativeness of the broader pollution field | Ranked candidate-site map with coverage radius estimates; GIS layer and written rationale |
| Diurnal NO2 cycle summary (GEMS coverage area) | Hourly GEMS NO2 retrieval aggregation over city extent; peak-to-trough ratio and hour-of-peak statistics | Time-series chart and summary statistics table; PDF briefing note |
Who does the work
We can get this done for you. Satellize runs its own analyst desk and a strong science team. You do not buy a data feed and work out what it means; our people source the imagery, run the analysis described on this page, and hand you the answer with its confidence limits stated. Discuss this requirement.